Tensor Field Networks for Rotation Equivariance in 3D Point Cloud Classification

TensorFlow implementation of Tensor Field Networks (https://arxiv.org/abs/1802.08219). Extended version of the code in https://github.com/tensorfieldnetworks/tensorfieldnetworks/tree/949e64ac6e069c2f1bfbcbf30d13f696a970488a. Batch learning is now supported. The proposed models are tested on ModelNet40 point cloud dataset (https://modelnet.cs.princeton.edu/). Developed and tested on Ubuntu 18.04 LTS.

Requirements

Anaconda Python >= 3.6.4 (see https://www.anaconda.com/distribution/);
pip (sudo apt install python3-pip to install it on Ubuntu 18.04 LTS);
virtualenv >= 16.6.0 (python3 -m pip install --user virtualenv to install it on Ubuntu 18.04 LTS).

Clone or download the repository and type the following commands in the root folder:

python3 -m venv env

source env/bin/activate

Now the virtual environment env is active (type deactivate if you want to deactivate it).

To install the dependencies, type the following command in the virtual environment:

pip install -r requirements.txt

Read modelnet/data/README.md for instructions on how to download ModelNet40 dataset.

python3 train.py to train the selected model. --help to show the help;
python3 evaluate.py to evaluate the selected model. --help to show the help;
read modelnet/tools/README.md for instructions on how to visualize the point clouds.